US Government Now Requires Pre-Release Safety Reviews for AI Models

What Are Pre-Release AI Reviews?

As of May 2026, the US government requires frontier AI models to undergo comprehensive safety, security, and capability assessments before public release. The Center for AI Standards and Innovation (CAISI) within NIST leads these evaluations, testing models for cybersecurity vulnerabilities, biosecurity risks, and other national security concerns.

Why This Shift Matters

This represents a fundamental change from voluntary oversight to mandatory federal evaluation. Major tech firms including Google, Microsoft, and xAI have agreed to provide early access to their frontier models for national security testing. CAISI has already completed over 40 evaluations of unreleased models.

Which Models Are Affected?

The reviews target "frontier AI models" — general-purpose models near cutting-edge performance whose training or capabilities present elevated national security risks. The threshold is defined by the National Security Memorandum NSM-25, issued in October 2024.

Impact on AI Development Timelines

Companies building frontier models must now factor review time into their release schedules. This could slow the pace of model launches but provides a safety net against potentially dangerous deployments.

Common Questions (FAQ)

Q1: Does this apply to all AI models? A1: No, only frontier models that meet specific capability and risk thresholds defined under NSM-25.

Q2: How long does a review take? A2: Timelines vary by model complexity. CAISI has completed 40+ reviews so far, suggesting a functional process.

Q3: What happens if a model fails review? A3: Models with significant safety concerns would need to address those issues before receiving clearance for public release.

Stay ahead of the AI curve. Follow @AiForSuccess for daily insights.